A figure steers a sailboat through an ocean of binary code.

AI’s Very Human Challenge

To ensure AI meets people’s needs and not vice-versa, we need data science—and humility.

By Scott L. Zeger • Illustration by Mark Smith

Novel technologies make us more efficient builders and destroyers. It’s a safe prediction that AI will be no different. 

I’m amazed each day by yet another AI news report. AI beats the world’s best Go player, who praises his opponent’s creativity. AI trains large language models on huge swaths of the internet. ChatGPT can’t directly count how many words in a paragraph, but it can write a computer program to do so, run the program, and give you the result.

Alan Turing, a founder of AI, imagined computing machines that were indistinguishable from humans. They far outstrip us now. Given that AI is the first technology to supersede human intelligence and may ultimately escape our control, we need to ensure AI meets human needs and not vice-versa.

For the past decade, I’ve been privileged to work with intelligent and wise clinicians. They have lots of experience and also something AI doesn’t (yet?) have: intuition. They can sense a patient’s health trajectory beyond what the recorded data might predict. In full partnership with clinicians, biostatisticians like me train statistical AI models that inform the care of a patient using the data from previous patients. But the physicians don’t turn the decision-making over to the machine. AI enhances their intelligence—it doesn’t replace it. To each patient, they bring all the knowledge, experience, and compassion they possess.  

Let me give you a concrete example. When COVID-19 cases surged at Johns Hopkins Hospital, I was part of a team of clinical and data scientists from Johns Hopkins Medicine, the Whiting School of Engineering, and the Bloomberg School. Our goal was to do data science and AI-powered, real-time analysis of COVID-19 patient data. The surge in cases in 2020 was overwhelming—as was the amount of patient data being captured. Doctors needed up-to-the-second guidance on which patients were at highest risk of intubation or death so they could triage their clinical efforts. Remember, this was a new disease with no effective treatment. We had to mobilize data on the fly, build prediction and evaluation models, and share the results with doctors on the wards. No human doctor could assimilate all the data on treatments and outcomes. With AI’s synthesis of that information, our clinicians could direct care more precisely.

We humans now face a challenge: to impose ethical and responsible control over AI. 

There were other COVID success stories operating at the population level. Data scientist Liz Stuart and her team combined data from diverse public sources to figure out the additional risk to children of returning to school. Nilanjan Chatterjee’s team designed interactive maps to guide public health agencies to areas where vaccine advocacy was most needed.

These are all examples of local, smart, effective use of AI. However, we also know AI can be used in dangerous ways—to power disinformation campaigns about vaccines or elections, and to concentrate power and wealth in ways that can harm the public.

We humans now face a challenge: to impose ethical and responsible control over AI. We must create policies to regulate AI, including testing for efficacy and safety, and longer-term monitoring and curation over the life of the product. When AI tools do harm, the owners and builders must be held responsible.  

How else do we keep the AI train from going off the rails? Data science, which overlaps with AI but has a distinctive focus, can help. It extracts insights to better understand an issue through the acquisition, organization, and analysis of complex data. Expertise in the given field is central to data science; less so to AI. A shining example is the Observational Health Data Sciences and Informatics (or OHDSI, pronounced “Odyssey”) program in which medical and statistical scientists around the world pool their medical data and collaborate in observational studies to better understand which treatments are best for which subsets of patients.

How might data scientists at Johns Hopkins and all over the world support AI to advance human health and other goals? First, we should heed the advice of legendary statistician John W. Tukey, a data science founder, who warned against statistical hubris—when we think we have more to offer to the solution of an important problem than we can actually deliver. Tukey prescribed daily doses of “anti-hubrisines” to prevent inflated promises, the way we take antihistamines to prevent symptoms of the common cold.

Human humility is the most important ingredient of a successful data science and AI partnership that accentuates AI’s positives and minimizes its risks.